The Significant Advantages to a Logically Segmented TBox

The Message Understanding Conferences (MUC) were initiated in 1987 and financed by DARPA to encourage the development of new and better methods of information extraction (IE). It was a seminal series that resulted in basic measures of retrieval and semantic efficacy, recall (R) and precision (P) and the combined F-measure, and other core terminology and constructs used by IE today.

By the sixth version in the series (MUC-6), in 1995, the task of recognition of named entities and coreference was added. That initial slate of named entities included the basic building blocks of person (PER), location (LOC), and organization (ORG); to these were added the numeric building blocks of time, percentage or quantity. The very terminology of named entity was coined for this seminal meeting, as was the idea of inline markup [1].

What is a ‘Nameable Thing’?

The intuition surrounding “named entity” and nameable “things” was that they were discrete and disjoint. A rock is not a person and is not a chemical or an event. As initially used, all “named entities” were distinct individuals. But, there also emerged the understanding that some classes of things could also be treated as more-or-less distinct nameable “things”: beetles are not the same as frogs and are not the same as rocks. While some of these “things” might be a true individual with a discrete name, such as Kermit the Frog, or The Rock at Northwestern University, most instances of such things are unnamed.

The “nameability” (or logical categorization) of things is perhaps best kept separate from other epistemological issues of distinguishing sets, collections, or classes from individuals, members or instances.

In a closed-world system it is easier to enforce clean distinctions. The Cyc knowledge base, for example, the basis for UMBEL (Upper Mapping and Binding Exchange Layer),  makes clear the distinction between individuals and collections. In the semantic Web and RDF, this can become smeared a bit with the favored terminology shifting to instances and classes, and in pragmatic, real-world terms we (as humans) readily distinguish John Smith as distinct from Jane Doe but don’t generally (unless we’re entomologists!) make such distinctions for individual beetles, let alone entire genera or species of beetles.

Under precise conditions, these distinctions are important. The fact that Cyc, for example, is assiduous in its application of these distinctions is a major reason for the overall coherence of its knowledge base. But, for most circumstances, we think it is OK to accept a distinction between “nameable” things such as frogs and beetles, but also to accept that there may be nameable individuals at times in those groupings such as Kermit that are truly an individual in that more refined sense.

This digression sets the background for a natural progression from that first MUC-6 conference. If we could cluster persons or organizations, why not other categories of distinct and disjoint things such as frogs or beetles or rocks?

From the first six entity categories of MUC-6 we begin to see an expansion to broader coverage. Readers of this blog will recall that I have been a fan for quite some time of the expanded coverage of 64 classes of entities proposed by BBN or the 200 proposed by Sekine [2] (as discussed, for example in the April 2008 Subject Concepts and Named Entities article). Again, the intuition was that real things in the real world could be logically categorized into discrete and disjoint categories.

Thus, “named entities” inexorably moved to become a categorization system, where the degree of familiarity and distinction dictated whether it was the individual (with a unique name, such as Abraham Lincoln or Mt. Rushmore) or groupings such as animal or plant species and their common names (such as beetle or oak) that was the standard “handle” for assigning a name to the “nameable thing”.

While many can argue these individual <–> grouping distinctions and whether we are talking about true, unique, named individuals or names of convenience, I think that (at least for this blog post and discussion), that misses the real, fundamental point.

The real, fundamental point is that some “things” (whether individuals, instances or classes) are distinct from other “things”. Such disjoint distinctions are a powerful concept that should not be lost sight of by “angels dancing on the head of a pin” epistemological arguments. A frog is not a rock, despite neither are “individuals”, and how can we take advantage of that realilty?

What Works for Entities, Works for Concepts

Nearly from the outset of our work with UMBEL as a ‘TBox’ [3] — that is, as a set of 20,000 or so common “subject concepts” — the natural question was what the relation or correspondence was of these concepts to the underlying “things” (entities) that they organized. As we probed the disjoint categories within the Sekine 200 entity types, for example, we began to see significant parallels and overlap. Also gnawing at our sense of order was the rather artificial and arbitrary class of concepts in UMBEL that we termed “Abstract Concepts”.

We introduced Abstract Concepts in the first release of UMBEL. When introduced, we defined “Abstract concepts [as] representing abstract or ephemeral notions such as truth, beauty, evil or justice, or [as] thought constructs useful to organizing or categorizing things but are not readily seen in the experiential world.” In pragmatic terms, Abstract Concepts in UMBEL were often pivotal nodes in the UMBEL subject graph necessary to maintain a high degree of concept interconnectivity.

In any world view that attempts to be more-or-less comprehensive, there is a gradation of concepts from the concrete and observable to the abstract and ephemeral. The recognition that some of these concepts may be more abstract, then, was not the issue. The issue was that there was no definable basis for segregating a concrete Subject Concept from the more Abstract Concept. Where was the bright line? What was the actionable distinction?

Off and on we have probed this question for more than a year, and have looked at what might constitute a more natural and logical ordering and segmentation within UMBEL. After many tests and detailed analysis, we are now releasing the first results of our investigations.

For, like nameable entities or things, we can see a logical segmentation of (mostly) disjoint concepts within the UMBEL TBox. Here are the summary percentages of these high-level splits:

(Because the analysis is still being refined, exact counts and percentages for the 20,000 concepts in UMBEL are not provided.)

Why a Logical Segmentation?

As we dove deeper into these ideas, not only could we see the basis for a logical segmentation within UMBEL’s concepts, but manifest benefits from doing so as well. Remember that UMBEL’s concept structure performs two main roles. It:  1) provides a coherent framework for relating and “mapping” other external ontologies; and 2) provides conceptual binding points for organizing entities and instances [4]. Via logical segmentation, we get benefits for both roles.

Here are some of the broad areas of benefit from a logical UMBEL segmentation that we have identified:

• Template-driven — as we discuss elsewhere, Structured Dynamics also uses its ontologies to “drive applications” and the user interfaces (UI) that support them. By proper segmentation of UMBEL concepts, we are able to determine to what “cluster” of things (which we call either dimensions or superTypes; see below) a given thing belongs. This identification means we can also determine how best to display information about that “thing”. This determination can include either the attributes or the display templates appropriate for that thing. For example, location-based things or time-based things might invoke map or calendar or timeline type displays. Moreover, because of the logical segmentation of concepts, we can also use the power of the concept graph to infer more generic display templates when specific matches are absent
• Computational Efficiency — as the percentages above indicate, once we identify what superType concept to which a given instance belongs, we can eliminate nearly all remaining UMBEL concepts from consideration. This logical winnowing leads to computational efficiencies at all levels in the system. The fastest computational work is not to do it, and when large chunks of data are removed from consideration, many performance advantages accrue
• Disambiguation — via this approach we now can assess concept matches in addition to entity matches. This means we can triangulate between the two assessments to aid disambiguation. Because of these logical segmentations, we also have multiple “clusters” (that is, either the concept, type, superType or dimension) upon which to do our disambiguation evaluations, either between concepts and entities or within the various concept clusters. We can do so via either multiple semantic vectors (for statistical-based methods) or multiple features (for machine learning methods). In other words, because of logical segmentation, we have increased the informational power of our concept graph
• Structure and Integrity Testing — the very mindset of looking for logical segmentation has led to much learning about the UMBEL structure and OpenCyc upon which it is based. In the process, missing nodes (concepts), erroneous assignments, and superfluous nodes are all being discovered. Further, many of these tests can be automated using basic logical and inference approaches. The net result is a constant improvement to the scope and completeness of the structure. Lastly, these same approaches can be applied when mapping external ontologies to UMBEL, providing similar consistency benefits.

With these benefits in mind, we have undertaken concerted analysis of UMBEL to discern what this “logical segmentation” might be. This investigation has occurred over three concentrated periods over the past year. (Intervening priorities or other work prevented concentrating solely on this task.)

We are now complete with our first full iteraton of investigation. In this post, and then the subsequent release of UMBEL version 0.80 in the coming weeks, the fruits of this effort should be evident. However, it should also be noted that we are still learning much from this new mindset and approach. UMBEL structure refinement may be likely for some time to come.

UMBEL Analysis

Most things and concepts about them are based on real, observable, physical things in the real world. Because most of these things can not occupy both the same moment in time and the same location in physical space, a useful criterion for looking at these things and concepts is disjointedness.

In a broad sense, then, we can split our concepts of the world between those ideas that are disjoint because they pertain to separable objects or ideas and those that are cross-cutting or organizational or classificatory. Attributes, such as color (pink, for example), are often cross-cutting in that they can be used to describe quite disparate things. Inherent classification schemes such as academic fields of study or library catalog systems — while useful ways to organize the world — are not themselves in-and-of the world or discrete from other ideas. Thus, classificatory or organizational concepts are inherently not disjoint.

With the criterion of disjointedness in hand, then, we began an evaluation process of the UMBEL subject concepts. We looked to organizational schema such as the entity types of Sekine or BBN for some starting guidance. We also kept in mind that we also wanted our categories to inform logical clusterings of possible data presentation, such as media types or locations or time.

For terminology, we adopted the term superType to denote the largest cluster designation upon which this disjointedness may occur. As a way to test the basic coherence of these superTypes, we also collected them into larger groups which we termed dimensions.

Our analysis process began with branch-by-branch testing of the UMBEL concept graph using automated scripts, attempting to find pivotal nodes where child instance members were disjoint from other superTypes. This we term the “top-down” method.

This automated analysis was then supplemented with a complete manual inspection of all unassigned and assigned concepts, with a “bottom up” assignment of concepts or corrections to the automated approach. This inspection then led to new insights and identification of missing concepts that needed to be added into UMBEL.

We are still converging between these two methods. Optimally, we should be able to tease out all UMBEL superTypes with a relatively few number of union, intersection, or complement set operations. In its current form, we are close, but there are still some rough spots.

Nonetheless, this analysis method has led us to identify some 33 superTypes [5], clustered into 9 dimensions. Of these, 29 superTypes and 8 dimensions are mostly disjoint. The one dimension of Classificatory includes the four cross-cutting superTypes of attributes and organizational schema that can apply to any of the 29 disjoint superTypes.

UMBEL superTypes

Here is the schema, with the descriptions of each:

These may undergo some further refinement prior to release of UMBEL v 0.80, and some of the definitions will be tightened up.

(Note: It should also be mentioned that some of these superTypes further lend themselves to further splits and analysis. The Product superType, for example, is ripe for such treatment.)

Distribution of superTypes

The following diagram shows the distribution of these 20,000 UMBEL concepts across major area. By far the largest superType is Products, even with further splits into Food and Drinks and Pharmaceuticals. The next largest categories are Person and Places and Events superTypes, with Organizations and Animals not far behind:

Even in its generic state, UMBEL provides a very rich vocabulary for describing things or for tying in more detailed external ontologies. There are nearly 5,000 concepts across products of all types, for example.

Possible Overlaps (non-disjoint) between superTypes

You may recall that our analysis showed 29 of the superTypes to be “mostly disjoint.”  This is because there are some concepts — say, MusicPerformingAgent — that can apply to either a person or a group (band or orchestra, for example). Thus, for this concept alone, we have a bit of overlap between the normally disjoint Person and Organization superTypes.

The following shows the resulting interaction matrix where there may be some overlap between superTypes:

This kind of interaction diagram is also useful for further analyzing the concept graph structure, as well.

Even Where Overlaps Occur, They are Minor

Of the 29 “mostly” disjoint superTypes, only a relatively few show potential interactions, and then only in minor ways. We can illustrate this (drawn to scale) for the interaction between the Product, Food & Drink and Drug (Pharmaceuticals) superTypes, with the fully disjoint Organization superType thrown in for comparison:

Across all 20,000 concepts, then, fully 85% are disjoint from one another (5% is lost due to overlaps between “mostly” disjoint superTypes). This is a surprising high percentage, with even better likelihood to deliver the benefits previously noted.

Interim Conclusions and Observations

These are exciting findings that bode well for UMBEL’s ongoing role and usefulness. Also, the very detailed analysis that has led to these interim findings very much reaffirms the wisdom of basing UMBEL on Cyc.  Cyc showed itself to be admirably coherent and remarkably complete. (It also appears that the first versions of UMBEL were also extracted well in terms of good coverage.)

This approach now gives us an understandable and defensible basis for logical segementation of UMBEL. It also provides a much-desired alternative to the earlier Abstract Concepts, which will now be dropped entirely as a schema concept.

One area deserving further attention is in the Attribute superType. We are in the process, for example, of analyzing attributes across Wikipedia and need to look through a slightly different lens at this superType [6]. This area is further important in its strong interaction with the Instance Record Vocabulary that is accompanying this effort on the entity side.

Another lesson for us has been to back away from the terminology of named entity, introduced at MUC-6. The expansions of that idea into other “nameable” things has caused us to embrace the “instance” nomenclature, as evidenced by our emerging IRV.

It is rewarding to prepare this next iteration release of UMBEL with its new mindset of logical segmentation and disjointedness. But — what is also clear — there are many treasures left to mine still hidden in the inherent structure of UMBEL and its Cyc parent.